Dataset · Kling-Keye
Benchmark for 4D scene generation in Blender by agents
Kling-Keye/vibeprevisbench
Test suite for evaluating AI agents' ability to write code for executable 4D scenes in Blender.
What's inside
A small benchmark of up to 1,000 English examples including text and video data. It tests whether coding agents can turn a description, script, photo, or initial frame into a ready animated Blender scene with object placement, motion, framing, and cuts. The project is accompanied by a website, an arXiv paper, GitHub code, and a leaderboard.
What you can do with it
- Evaluating AI agents in code-based 3D/4D animation generation tasks
- Testing multimodal models on understanding direction, framing, and editing
- Comparing automated scene pre-visualization systems based on scripts or references
- Tasks
- Text → 3d, Image → 3d
- Languages
- English
- Size
- up to 1K rows
- Downloads this month
- 2.8K
- Likes
- 4
How to load it
from datasets import load_dataset
ds = load_dataset("Kling-Keye/vibeprevisbench")